YOLO-AKECA: Enhancing Pediatric Wrist Fracture Detection with Alterable Kernel Convolution and Efficient Channel Attention in YOLOv9

Abstract

This research endeavors to optimize the YOLOv9 algorithm's efficacy in detecting pediatric wrist fractures in X ray imagery by incorporating the Alterable Kernel Convolution (AKConv) and Efficient Channel Attention (ECA) techniques. While the YOLOv9 model has proven to be a robust object detection tool in a multitude of sectors, its performance in handling low-detail X-ray images can be further refined. Our innovative YOLO-AKECA framework markedly boosts the delineation of features via the ECA component and more adeptly accommodates variations in the shape of targets with the AKConv feature. Empirical data from the GRAZPEDWRI-DX dataset indicates an enhancement of 3.7 percentage points in the crucial metric mAP for our advanced model. Moreover, comprehensive performance evaluations against other models and meticulous ablation analyses were executed to substantiate the effectiveness of our strategy. This inquiry not only extends the utility of YOLOv9 in medical imaging but also paves the way for forthcoming explorations in the realm of intricate image datasets.

Ähnliche Arbeiten